GraphRAG

Standard RAG (retrieval-augmented generation) works by turning text into vectors (numeric representations) and finding the chunks closest to your query. That works well when the answer lives in a single passage. It struggles when the question is relational: how is concept A connected to concept B across fifty documents, or what are the main themes across an entire corpus? GraphRAG was introduced by Microsoft Research to address exactly those gaps.

The approach has two phases. In the indexing phase, an LLM reads your documents and extracts entities (people, products, concepts, places) and the relationships between them, assembling everything into a knowledge graph. Groups of closely related entities get clustered and summarized. In the retrieval phase, a query can traverse the graph along those relationships rather than just pulling the nearest vector matches, enabling multi-hop reasoning and global summarization that vector-only retrieval misses.

GraphRAG is meaningfully harder and more expensive to set up than standard vector RAG. Indexing requires running an LLM over the entire document corpus, which costs money and time. For many use cases, plain RAG is good enough. But in domains where relationships between entities matter, like legal research, competitive intelligence, or enterprise knowledge management, the extra fidelity is often worth it. Microsoft's open-source implementation has been one of the most-starred RAG repositories on GitHub through 2026.